What is harness uplift?

Two submissions with the same LLM can score 5 to 30 points apart on SWE-bench Verified depending only on the harness around the model — the tool loop, the context-management policy, the planner. agent-uplift surfaces that delta for every frontier model. Pick a model, see the bare-baseline score, the best-known ceiling, and the harness that gets you there.

All data is pulled hourly from swe-bench/experiments. No mock data. No seeded fallback. Methodology.